Hyperparameter search

.github/workflows/ci-tests.yaml

@@ -10,6 +10,7 @@ jobs:
   # tests the libraries that our projects depend on
   library-ci-tests:
     runs-on: ubuntu-latest
+    if: github.event.pull_request.draft == false
     permissions:
       packages: read
     container:

libs/trainer/deepclean/trainer/trainer.py

@@ -285,7 +285,7 @@ def train(
         batch_size=batch_size,
         chunk_length=chunk_length,
         num_chunks=num_chunks,
-        shuffle=False,
+        shuffle=True,
         device=device,
     )
 
@@ -423,6 +423,7 @@ def train(
 
     # generate some analyses of our model
     logging.info("Performing post-hoc analysis on trained model")
+    train_data.batch_size = 32
     gradients, coherences = analyze_model(train_data, model, sample_rate)
     history.update(
         {
@@ -452,4 +453,4 @@ def train(
     nn = PrePostDeepClean(model)
     nn.fit(X, y)
     torch.save(nn.state_dict(), weights_path)
-    # return history
+    return history

projects/sandbox/pyproject.toml

@@ -95,6 +95,45 @@ verbose = "${base.verbose}"
 # architecture subcommands
 commands.autoencoder = "${base.autoencoder}"
 
+[tool.typeo.scripts.search]
+output_directory = "${PROJECT_DIR}"
+data_path = "${DATA_DIR}"
+
+# data parameters
+channels = "${base.channels}"
+t0 = "${base.train_t0}"
+duration = "${base.train_duration}"
+valid_frac = 0.25
+
+# data loading parameters
+sample_rate = "${base.sample_rate}"
+kernel_length = 8
+kernel_stride = 0.25
+chunk_length = 0
+
+# preprocessing parameters
+freq_low = "${base.freq_low}"
+freq_high = "${base.freq_high}"
+filter_order = 8
+
+# optimization parameters
+batch_size = 64
+max_epochs = 500
+num_trials = 10
+patience = 8
+factor = 0.2
+early_stop = 50
+
+# criterion parameters
+fftlength = 2
+overlap = 1
+alpha = 1.0
+
+# miscellaneous parameters
+profile = false
+verbose = "${base.verbose}"
+
+
 [tool.typeo.scripts.export-model]
 repository_directory = "${MODEL_REPOSITORY}"
 output_directory = "${PROJECT_DIR}"